Abstract
Longitudinal imaging studies are essential to understanding the neural development of neuropsychiatric disorders, substance use disorders, and the normal brain. This paper develops a two-stage adjusted exponentially tilted empirical likelihood (TAETEL) for the spatial analysis of neuroimaging data from longitudinal studies. The TAETEL method allows for the efficient analysis of longitudinal data without correctly modeling temporal correlation and classifies different time-dependent covariate types. To account for spatial dependence, TAETEL combines data in the neighborhood of each voxel with appropriate weights to calculate adaptive parameter estimates and test statistics. We demonstrate the application of our statistical methods to the detection of morphological changes in the hippocampus over time between schizophrenia patients and healthy subjects.
Introduction
Tracking brain changes over time through longitudinal neuroimaging is critical for understanding the progression of disorders like schizophrenia. However, this data presents significant statistical challenges. Scans from the same individual are naturally related (temporally correlated), and failing to model this relationship accurately in standard parametric models can lead to biased or misleading results. Furthermore, time-dependent covariates, such as Whole Brain Volume (WBV), fluctuate alongside the specific regions of interest, complicating the isolation of localized effects. Finally, traditional spatial smoothing techniques often apply a uniform "blur" that can obscure sharp anatomical boundaries and mix signals from distinct tissues.
Methodology: The TAETEL Framework
The proposed Two-stage adjusted Exponentially Tilted Empirical Likelihood (TAETEL) framework addresses these challenges through a robust non-parametric approach.
Stage 1: Voxel-wise Analysis
Instead of assuming a Gaussian distribution, TAETEL utilizes Empirical Likelihood (EL), a semi-parametric method that allows the data distribution to be determined empirically. An adjusted version (AETEL) is employed to maintain robustness in the presence of small sample sizes or outliers. In the first stage, the model is fitted independently at every voxel to obtain initial parameter estimates.
Stage 2: Adaptive Spatial Smoothing
The second stage introduces "adaptive" smoothing. Unlike standard Gaussian kernels that blur everything equally, TAETEL assigns weights to neighboring voxels based on the similarity of their signal patterns identified in Stage 1. Neighbors with similar patterns receive high weights, while those across structural boundaries receive low weights. This edge-preserving mechanism enhances statistical power without the typical cost of losing spatial resolution at the borders of brain structures.
Results and Discussion
The performance of TAETEL was evaluated using both simulations and a real-world longitudinal schizophrenia study involving 238 patients and 56 healthy controls, with MRI scans taken at baseline and up to seven follow-up points.
Statistical Efficiency and Sensitivity
Simulation studies revealed that TAETEL achieved a 1.82x efficiency gain over standard Generalized Estimating Equations (GEE) when handling complex time-dependent covariates. In the clinical dataset, the method proved more sensitive than baseline comparisons, identifying significant morphological differences in the superior and anterior regions of the left hippocampus, specifically near the cornu ammonis 1 (CA1) and 2 (CA2) subfields.
Robustness to Non-Gaussian Data
Shapiro-Wilk tests confirmed that the error distributions in the neuroimaging data were non-Gaussian. This finding justifies the use of TAETEL’s non-parametric approach over traditional software packages like SPM (Statistical Parametric Mapping), which rely on rigid parametric assumptions.
Conclusion
TAETEL provides a robust and efficient framework for analyzing longitudinal neuroimaging data. By categorizing time-dependent variables and employing adaptive spatial smoothing, it offers superior detection of localized brain changes while preserving anatomical integrity. Future work will focus on reducing computational intensity and integrating Random Field Theory (RFT) specifically tailored for this non-parametric model to better control for whole-brain multiple comparisons.
Summary updated