A Bayesian Multilevel Dirichlet Regression Model for Adolescent Activity Pattern Data

• Activity Number: 249 - Bayesian Methods for Social and Human Data
• Type: Contributed
• Date/Time: Monday, July 30, 2018 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Kori Khan* and Catherine Calder and Anna Smith and Christopher Browning
• Companies: Ohio State and The Ohio State University and Columbia University and The Ohio State University
• Keywords: Bayesian statistics; Hierarchical modeling; Neighborhood effects; Social science

Abstract

Our analyses explore the link between the socioeconomic characteristics of residential neighborhoods and the allocation of adolescent waking time to home, neighborhood, and outside-neighborhood spaces for urban youth. We use novel space-time data collected using Smartphone-based GPS from the Adolescent Health and Development in Context (AHDC). The data capture activity trajectories over the course of a five-day period for 1400 adolescents ages 11-17 residing in an urbanized area in and around Columbus, OH. Using a reparameterization of the Dirichlet distribution that allows us to employ a mean and dispersion like model, we fit a multi-level Bayesian Dirichlet regression model relating the expected proportion of time in each time-use category to covariate information at the day, individual, and neighborhood levels. We find that youth from residential neighborhoods with high socioeconomic disadvantage spend a lower proportion of time in neighborhood and a higher proportion of time out of neighborhood relative to youth in residential neighborhoods with low socioeconomic disadvantage.