Bayesian multivariate longitudinal piecewise regression: detecting early onset of cognitive decline

The current paper defines and tests a hierarchical multivariate change point model (HMLCP) to detect a random change point in longitudinal data (subject-specific). Given the large number of random effects that are estimated, we implemented the HMLCP using Bayesian methods. First, we present model parameterization and then applied it in a sample of older adults, to identify the onset of cognitive decline with neurocognitive data collected at the Knight Alzheimer Disease Center. These data presented an ideal case study to test a multivariate random change-point model with a diagnosed sample, to see if the model can detect the disease progression of AD. HMLCP model was sensitive to detect the presence of disease several years before clinical diagnosis. Further, the HMLCP allowed us to describe the sample disease progression, and predict change points at the level of the individual. We demonstrate HMLCP’s flexibility by comparing results from the whole sample with individual subjects and conclude that the HMLCP is a powerful model for estimating multivariate-multilevel unknown change point trajectories.