Co-location networks ? two-mode networks that capture connections between individuals and locations in geographic space ? have broad relevance in the health sciences in areas ranging from the study of infectious disease transmission to understanding the influence of social processes on health outcomes and behaviors. Despite their broad relevance, however, statistical methods for understanding co-location networks are limited. This methodologically oriented proposal focuses on the development of a statistical framework for the study of co-location networks using a bilinear mixed-effects model with interacting latent activity pattern motifs and profiles. Through latent interacting random effects, our model captures the dependence between individuals based on their shared use of space and between locations based on the individuals who frequent them. Our flexible modeling framework uses a mixed-membership structure to relax the assumption that activity profiles are static and takes advantage of a data augmentation strategy to allow versions of the model with either direct or indirect specification of the dependence between actor-location ties. Our novel statistical models will be used in analyses of activity pattern data collected as part of the Adolescent Health and Development in Context (AHDC) Study, an ongoing data collection effort in Franklin County, Ohio. Through GPS-based smartphone tracking and space-time budget software, the AHDC Study provides rich detail on the co-location networks of adolescents in the study area. In addition, a wealth of survey data, smartphone-administered Ecological Momentary Assessments (capturing real-time measures of location, social network partner presence, activities, risk behaviors, and mood), and biomeasure data on the study participants are available. Recognizing that our proposed statistical model may not be able to capture the structure the co-location network structure of AHDC adolescents based entirely on their observed activity patterns, we also propose to embed relevant information derived from social media into our analyses through informative prior distributions on model parameters. To do so, we propose novel data mining algorithms to retrieve potential activity pattern motifs and coincident profiles from Twitter posts and network structure. In particular, we extend named entity identification methods to the spatial setting to automatically retrieve information relevant to activity patterns and develop novel methods for prioritizing activity pattern information based on its relevance to particular subpopulations (here, adolescents) using scalable sentiment analysis. Using our new statistical and data mining methodology, we will perform detailed statistical analyses to explore the relationship between spatial and socio-spatial exposures derived from an inferred co-location network and physiological stress in adolescents.
There is increasing recognition that the area around an individual's residence is not the only spatial setting relevant for capturing health consequential environmental and social exposures ? individuals interact with the environment and others as they spend time away from home performing their routine activities (e.g., work, school, shopping). Findings from this research will allow a better understanding of how spending time in violent-prone locations and areas with high levels of social/physical disorder, in addition to routinely encountering individuals who spend time in such areas, affect the levels of physiological stress in adolescents residing in an urban settings.
