Bayesian Methods for Individualized Trajectory Models
Abstract: A trajectory of observations from a single unit has become increasingly commonplace in modern scientific studies. Examples include trajectories of patient observations (made easier by state of the art collection of EHR data), tracking the motion of an object over time, tracking changes in tissue or deformations, among others. Traditional statistical techniques rely on the use of some modification of mixed models, where the repeated component inside the mixed model is used to accommodate the “trajectory”. This is however untenable in many application areas – simple because assuming a uniform trajectory for the entire population under consideration is unrealistic. The other extreme would be to assume each individual unit has their own trajectory formulation, which is prohibitive in terms of statistical cost – the number of parameters being modelled. I will demonstrate the use of a semiparametric Bayesian model driven soft clustering for trajectories, which can be embedded inside the traditional mixed models, resulting in a very flexible estimation structure. Under this framework, I will discuss estimation strategies and model parsimony using a plethora of examples, from high frequency observations in biomechanical contexts, to relatively low frequency observations in EHR data.