Multivariate finite mixture latent trajectory models with application to dementia studies

Abstract

Dementia studies often collect multiple longitudinal neuropsychological measures in order to examine patients' decline across a number of cognitive domains. Dementia patients have shown considerable heterogeneities in individual trajectories of cognitive decline, with some patients showing rapid decline following diagnoses while others exhibiting slower decline or remain stable for several years. In the first part of this dissertation, a multivariate finite mixture latent trajectory model was proposed to identify longitudinal patterns of cognitive decline in multiple cognitive domains with multiple tests within each domain. The expectation-maximization (EM) algorithm was implemented for parameter estimation and posterior probabilities were estimated based on the model to predict latent class membership. Simulation studies demonstrated satisfactory performance of the proposed approach. In the second part, a simulation study was performed to compare the performance of information-based criteria on the selection of the number of latent classes. Commonly used model selection criteria including the Akaike information criterion (AIC), Bayesian information criterion (BIC), as well as consistent AIC (CAIC), sample adjusted BIC (SABIC) and the integrated classification likelihood criteria (ICLBIC) were included in the comparison. SABIC performed uniformly better in all simulation scenarios and hence was the preferred criterion for our proposed model. In the third part of the dissertation, the multivariate finite mixture latent trajectory model was extended to situations where the true latent class membership was known for a subset of patients. The proposed models were used to analyze data from the Uniform Data Set (UDS) collected from Alzheimer's Disease Centers across the country to identify various cognitive decline patterns among patients with dementia.