UCLA

It is often of interest to study the temporal patterns of longitudinal biomarker(s) that are potentially correlated and predictive of time-to-event outcomes in biomedical studies. In this dissertation, I develop new approaches for joint modeling of longitudinal and time-to-event data. My dissertation consists of three projects. In Chapter 2, I develop customized linear scan algorithms to speed up the computation of semi-parametric joint models by linearizing the computational burden of the estimation procedure from $O(n^2)$ or $O(n^3)$ to $O(n)$. Compared to the existing software and packages on semi-parametric joint models, our implementations can provide more than thousands of speed-ups when the sample size goes large. In Chapter 3, motivated by the Multi-Ethnic Study of Atherosclerosis (MESA), I propose a novel joint model to account for the heterogeneity of within-subject variability of a longitudinal outcome and demonstrate that it improves the dynamic prediction accuracy of predicting the future event probabilities of both heart failure and death across MESA individuals. In Chapter 4, I extend the joint model described in Chapter 3 to handle interval-censored covariates as missing data due to the unknown initial event time. Using age at diagnosis of diabetes as an interval-censored covariate, we revisit the UK-Biobank data to illustrate that our proposed joint model can yield clinically meaningful parameter estimates, compared to the existing methods such as midpoint imputation, which can lead to problematic conclusion on the effect of covariates on the outcomes.