Title: Modeling flexible trajectories and related outcomes using a three-level enriched Dirichlet process mixture
Authors: Natalie Burns - University of Florida (United States) [presenting]
Michael Daniels - University of Florida (United States)
Elizabeth Widen - University of Texas at Austin (United States)
Abstract: Dirichlet process mixture (DPM) models can be used for density estimation and clustering. When jointly modeling a response and covariates, the enriched Dirichlet process mixture (EDPM) overcomes the possibility that the covariate measurements will dominate the response measurements in the clustering structure induced by the DPM. A further extension of the EDPM is proposed, the three-level EDPM (EDP3) with flexible trajectories. In the motivating example, there are three sets of variables: many measurements of gestational weight gain (GWG); several neonatal size outcomes; and a number of other maternal variables. The EDP3 induces a three-level nested clustering structure on the data. Nesting the clustering of the neonatal outcomes within the top-level GWG trajectory clusters, combined with relabeling of the MCMC output to address the label-switching problem, allows us to make meaningful interpretations of the top-level components to identify various GWG trajectory classes, which characterize different patterns of weight gain, and analyze the distributions of the neonatal outcomes within each GWG trajectory class. Further, including the maternal covariates in the third level of clustering ensures that the random partitions of subjects within each GWG trajectory class rely sufficiently on the neonatal outcomes and are not dominated by the maternal characteristics.