A1177
Title: Model selection for Bayesian dynamic clustering factor models
Authors: Tsering Dolkar - Virginia Tech (United States) [presenting]
Marco Ferreira - Virginia Tech (United States)
Allison Tegge - Virginia Tech (United States)
Abstract: The purpose is to consider model selection for Bayesian dynamic clustering factor models (BDCFMs). BDCFMs are novel models that combine latent factor models and hidden Markov models (HMMs) for the analysis of multivariate longitudinal data. As such, BDCFMs perform concomitant dimension reduction, clustering, and estimation of the dynamic transitions of subjects among clusters. To select the number of clusters and the number of factors in BDCFM, an information criterion inspired by the Bayesian information criterion is proposed. To implement the information criterion, the computation of the likelihood for HMMs is extended using a forward-backward recursion to BDCFMs. A simulation study shows that, when compared to competing approaches, the approach has favorable performance. The utility of the approach is shown with an application to a longitudinal study on recovery from substance use disorder.
