Title: Finite mixture models for longitudinal data with dynamic group membership
Authors: Jeffrey Andrews - University of British Columbia Okanagan (Canada) [presenting]
Liam Welsh - University of Toronto (Canada)
Ryan Browne - University of Waterloo (Canada)
Abstract: A compositional approach is introduced for the building and fitting of a finite Gaussian mixture model, permitting highly constrained components to be added to the model at a very low cost with respect to growth in free parameters. The explicit goal of this approach is to enable both the detection and modelling of small numbers of observations which change groups over time in longitudinal data --- all under a fully unsupervised paradigm. The proposed approach can be considered an alternative to others in the literature which rely on hidden Markov models to achieve a similar effect. We provide both simulations and real data applications for illustrative purposes.