A LATENT CLASS APPROACH FOR JOINT MODELING OF CLINICAL OUTCOMES AND LONGITUDINAL BIOMARKERS SUBJECT TO DETECTION LIMITS

Restricted (Penn State Only)
Author:
Li, Menghan
Graduate Program:
Biostatistics
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
April 30, 2018
Committee Members:
  • Lan Kong, Dissertation Advisor
  • Lan Kong, Committee Chair
  • Vernon Michael Chinchilli, Committee Member
  • David Theodore Mauger, Committee Member
  • Xuemei Huang, Outside Member
Keywords:
  • JOINT MODELING;
  • MCEM;
  • BIOMARKER
Abstract:
Biomarkers have played an important role in the biomedical studies for diagnosis and prognosis of acute and chronic diseases, and understanding the biological mechanisms of disease development and treatment effects. Joint modeling of longitudinal biomarkers and clinical outcome provides an effective method to study the associations between biomarkers and disease progression. This research is motivated by the Genetic and Inflammatory Markers of Sepsis (GenIMS) study, where multiple biomarkers were collected during the course of hospitalization in patients admitted to the emergency department with community acquired pneumonia. One of the primary goals of GenIMS study is to investigate the systemic cytokine response to pneumonia and determine if specific patterns are associated with the risks of severe sepsis and death. In this dissertation, we focus on joint modeling the trajectories of biomarkers and clinical outcomes of interest using a latent class approach. Joint latent class models assume a heterogeneous population and characterize the associations between biomarkers and clinical outcomes via a latent class structure. This modeling approach can be used to identify informative subgroup patterns of biomarker profiles and a clinical endpoint, and predict the risk of an event of interest given longitudinal biomarker information and other clinical or demographic characteristics. There are two challenges in the analysis of GenIMS data. Due to the low sensitivity of bioassays, several cytokine biomarkers are censored at lower detection limits. Simple substitution of censored data with detection limits may lead to a substantial bias in model estimation. Furthermore, the interrelationship between multiple biomarkers needs to be taken into account in joint latent class models. This dissertation aims to provide appropriate and efficient statistical methods to handle censored biomarker data in joint latent class models. Under the parametric distributional assumption of biomarkers, we developed likelihood based approaches to account for censored biomarker measurements in joint latent class models. In specific, we first proposed a joint model for one longitudinal biomarker and a binary outcome. Then we extended our model to incorporate a survival outcome. Next we laid out the framework for handling multiple censored biomarkers. The estimation of joint latent class models becomes more complicated when biomarkers are subject to detection limits. We presented novel Monte Carlo Expectation-Maximization algorithms to obtain the maximum likelihood estimates of model parameters, and evaluated the estimation procedures via comprehensive simulation studies. For demonstration, we applied our proposed methods to the GenIMS study to examine the heterogeneous patterns of cytokine responses and mortality risks in patients with pneumonia.