Identifying brain-behavior associations for the purpose of informing individual differences, illness trajectories, and neural mechanisms is one of the primary goals of psychiatric neuroimaging. The massively multivariate nature of neuroimaging data, which consists of spatially detailed images of brain structure and function, combined with high-dimensional behavioral data pose significant challenges to meeting this goal. The emerging replication crisis in neuroimaging research has exposed limitations of commonly used spatial extent inference (SEI) methods for analyzing imaging data. These include unrealistic assumptions about the spatial covariance function of the imaging data that lead to highly inflated error rates. This project will develop a new robust semiparametric inference framework for neuroimages to address the need for methods that are robust in real-world data, integrate these methods into the pbj R package, and develop a graphical user interface (GUI) to make the methods accessible to neuroimaging scientists. We will use the methods to study how multidimensional symptoms of psychosis are related to brain function and structure in the Psychiatric Genotype-Phenotype Project (PGPP) collected and Vanderbilt University Psychiatric Hospital (VUPH) and to study cross-sectional and longitudinal changes in functional connectivity in the public-access Nathan Kline Institute Rockland Sample (NKI-RS). We will evaluate the methods using realistic bootstrap-based neuroimaging simulations.
In Aim 1 we will develop a multidimensional semiparametric procedure for SEI that will leverage computationally efficient parametric and nonparametric bootstraps for inference.
In Aim 2 we will expand the framework to repeated measurement models (including longitudinal data), that will allow scientists to robustly model associations of subject-level covariate measurements and brain structure or function.
In Aim 3, to address the need for alternatives to hypothesis testing in psychiatric neuroimaging, we will develop semiparametric Coverage Probability Excursion (CoPE) sets that can be used to construct spatial confidence intervals for semiparametric effect sizes. These methods will be made available to the neuroimaging community through the pbj R package and GUI, and disseminated at neuroimaging conferences.
An important goal in psychiatric neuroimaging is to identify brain-behavior associations for the purpose of identifying biomarkers that are indicative of illness trajectory or inform our biological understanding of psychiatric disorders. The goal of this project is to develop a statistically robust framework to address the limitations of existing inference methods, which have highly inflated false positive rates in many situations. This study will impact public health by improving the rigor and replicability of statistical inference for neuroimaging studies so that results can more accurately inform our understanding of the biological underpinnings of psychiatric illness.
