Modifiable behaviors are a leading cause of preventable death, and decades of work have examined how to motivate change. Prior research has offered theoretical arguments and empirical evidence in support of many factors contributing to (un)healthy behaviors, yet much is still unknown about which of many factors matter most, under what circumstances, and for whom. This information is key to making precise predictions about future behavior and to design optimal behavior change interventions. This dissertation employed a person-specific (N-of-1) approach to predict two common health behaviors, alcohol use, and physical activity at the individual level among young adults. We applied machine learning to twice-a-day smartphone sampling data over 28 days and compared personalized models—which assume that factors contributing to behavior are idiosyncratic—to aggregate (cold start) models—which assume shared factors. In Study 1, we drew on affect and social context factors to predict future alcohol episodes across two samples (N=46, 2484; N=41, 2214 observations). In Study 2, we assessed the feasibility of the person-specific approach to predict physical activity episodes, while also integrating geospatial factors (N=64, 3456 observations). In both studies, we found that personalized models led to some improvement in predictability relative to aggregate models, in cases where data was available. However, individuals varied substantially in their predictability, and many individuals were difficult to predict. In Study 3, we applied the person-specific approach to evaluate competing perspectives on the role of affective factors (N=41, 2484 & N=64, 3456). Here, we found that discrete emotions provided little improvement beyond general valence to predict future alcohol use and physical activity. Taken together, across two behavioral contexts, we find that different factors matter to predict behavior for different individuals, and accounting for this heterogeneity can, in some cases, improve predictability. Our work suggests a predictive benefit of certain kinds of personalization, where data is available, and highlights an opportunity to triangulate between aggregate and personalized models for improved scalability. Overall, this work supports the feasibility of the person-specific approach and advocates for characterizing inter-individual heterogeneity to build more predictive theories of health behavior and optimal alcohol and physical activity interventions.