Date of Award
Dissertation - MUSC Only
Doctor of Philosophy (PhD)
Public Health Sciences
College of Graduate Studies
The observational study of subjects with neurological disease gives rise to complex data. This dissertation, consisting of three aims, concerns the development of statistical models to address these complex data. Aim one of this dissertation concerns the complex data scenario where we wish to model counts of multiple diseases over space and time that rep- resent progressive stages of neurodegeneration. Counts of multiple diseases over space and time are likely to be correlated. Therefore an appropriate model for these disease count data should account for this correlation. We thus propose a novel Bayesian space-time disease count model which can account for the correlation between multiple diseases. We demonstrate that this novel model provides superior prediction of counts of multiple diseases over space and time. Aim two of this dissertation concerns the complex data scenario where we wish to identify subject feature data that are associated with longitudinal outcome data of subject cognition. These longitudinal outcome data are irregularly-spaced over time, these feature data may be time-invariant, and these feature data may have group structure. Therefore an appropriate feature selection model should specify a covariance structure for these longitudinal outcome data which is a function of time separation, should specify time-varying feature parameters, and should account for group structure in these feature data. We thus propose a novel Bayesian feature selection model which can account for these complexities. We demonstrate that this novel model provides excellent feature selection performance as well as superior prediction of irregularly-spaced longitudinal outcome data. Novel statistical models are underused due in part to a lack of dissemination. Therefore, aim three of this dissertation concerns the development of an R software package that implements our Bayesian feature selection model from aim two. This software package promotes a rigorous and reproducible model-fitting workflow, facilitates the seamless summarization and visualization of clinically-relevant model results, and disseminates our feature selection model for use in the analysis of complex data e.g. stemming from the observational study of neurological disease.
Baer, Daniel, "The Development of a Space-Time Mixture Model and a Bi-Level Feature Selection Model with Applications to Neurological Data" (2021). MUSC Theses and Dissertations. 583.
All rights reserved. Copyright is held by the author.