Author ORCID Identifier
0000-0002-5050-443X
Date of Graduation
8-2024
Document Type
Dissertation (PhD)
Program Affiliation
Biostatistics, Bioinformatics and Systems Biology
Degree Name
Doctor of Philosophy (PhD)
Advisor/Committee Chair
Liang Li
Committee Member
Jing Ning
Committee Member
Xuelin Huang
Committee Member
Peng Wei
Committee Member
Brad Astor
Committee Member
David Ost
Abstract
Dynamic prediction plays a pivotal role in clinical research, especially when forecasting time-to-event outcomes based on evolving longitudinal data. This process often leverages the integration of longitudinal and time-to-event data through joint modeling, a prevalent technique. Alongside joint modeling, landmark modeling stands as another key approach in the realm of longitudinal studies. These methodologies are instrumental in dynamically predicting clinical events by utilizing predictor variables measured over time, up until the moment predictions are made. Within this framework, Chapter 2 addresses the challenge of comparing joint modeling and landmark modeling for dynamic prediction in longitudinal studies, introducing a novel algorithm that generates data aligning with landmark model assumptions for a balanced evaluation. The study reveals that neither approach consistently outperforms the other in prediction accuracy, emphasizing the need for methodological advancements in both areas, illustrated through a kidney transplantation dataset analysis. Chapter 3 introduces a novel joint model for dynamically predicting end-stage renal disease with the competing risk of death, motivated by a study on chronic kidney disease, which efficiently handles the computational challenges of using multiple longitudinal biomarkers through a different factorized likelihood approach and EM algorithm. The model, which facilitates the prediction of future longitudinal data trajectories conditional on future risk, demonstrates improved efficiency and stability over conventional shared random effects models, validated through simulations and a real dataset comparison. Chapter 4 introduces the multi-layer backward joint model (MBJM) for dynamic prediction in clinical research, designed to efficiently handle multivariate longitudinal predictors of both continuous and categorical types, overcoming the computational limitations of traditional shared random effect models. The MBJM, validated through comparative analysis and application to the Mayo Clinic's Primary Biliary Cirrhosis dataset, demonstrates superior predictive accuracy and computational efficiency, highlighting its potential as an advanced tool for complex medical datasets.
Keywords
Dynamic prediction, Joint model, Competing risks, Multivariate longitudinal data, Survival analysis, Landmark analysis
Included in
Biostatistics Commons, Longitudinal Data Analysis and Time Series Commons, Survival Analysis Commons