Advisory Committee Chair
Inmaculada Aban
Advisory Committee Members
Hemant Tiwari
Leslie A McClure
Louis Dell'Italia
Thomas Denney
Himanshu Gupta
Document Type
Dissertation
Date of Award
2014
Degree Name by School
Doctor of Philosophy (PhD) School of Public Health
Abstract
Longitudinal imaging studies have increased in popularity as clinical researchers seek to investigate how phenomena within the body change over time. Analysis of data from these studies is complicated by correlation between repeated measures over time and different locations in the body. To address this problem we propose the use of a linear model with a separable parametric correlation structure. This model considers spatial and temporal correlation independently and incorporates the correlation using parametric functions that have the potential to be much more efficient than an unstructured approach. Our model also has the ability to control for time- and space-varying covariates, which previously used summary methods cannot do. Results from a simulation study that investigates the effects of correlation structure selection on statistical inference about a treatment-by-time interaction are reported. Using the true correlation structure conserves the Type I error rate and maximizes power versus other structures, while misspecified structures may inflate the type I error rate or reduce power. If the misspecified structure can closely approximate the true correlation function then the Type I error rate is conserved and the loss of power from the true structure is negligible. For the considered conditions, information criteria are highly accurate at choosing a working correlation structure that conserved the Type I error rate. Our model is compared to summary methods through a simulation study that considers inference on a treatment-by-time effect. Our model more reliably conserves the Type I error rate and has greater statistical power than summary methods in space and time. The practice of analyzing spatial regions separately is found to have poor statistical properties. The presence of missing data does not change the qualitative results. Finally, we apply our model to the UAB SCCOR study, which considered MRI-derived outcomes from a longitudinal clinical trial in mitral regurgitation patients assigned to medical therapy or placebo. This study provided the motivation for this dissertation and inspired the scenarios used in the simulation studies. Here we discuss practical considerations of applying our model to real data such as how to choose a working correlation structure and how to handle missing data.
Recommended Citation
George, Brandon, "A Spatiotemporal Model for Repeated Imaging Data" (2014). All ETDs from UAB. 1728.
https://digitalcommons.library.uab.edu/etd-collection/1728
