Clinical neuroscience research using functional magnetic resonance imaging (fMRI) to inform individualized diagnosis and therapy is poised to flourish under major initiatives, but more accurate fMRI statistical methodology is needed. This proposed research will address open questions regarding the reliability and validity of major fMRI statistical methods and lead to the development of better-powered multilevel fMRI methods.
The first aim addresses test-retest reliability and validity of fMRI network-mapping. Standards for fMRI functional network-mapping have been proposed based on test-retest reliability. Previous studies have offered insight into test-retest reliability of functional connectivity, but due to differing analytical approaches and focuses, contradictions have risen and open questions remain. These include: (1) how amount of data per subject influences whole-brain reliability, (2) the influence of increasing runs versus sessions, (3) the spatial distribution of reliability, (4) the reliability of multivariate methods, and, crucially, (5) how reliability maps onto utility (prediction of behavior). My completed dissertation research, recently published, addressed these questions using the Generalizability Theory framework for assessing test-retest reliability and an extensively sampled dataset.
The second aim addresses validity of fMRI activation-mapping. Recent work demonstrated that conventional parametric fMRI activation-mapping analyses result in insufficient specificity. Namely, the familywise error rate (FWER) did not reach expected levels. However, that study's emphasis on specificity, although important, is incomplete. The authors found non-uniform specificity at the level of individual tests, marked by a greater incidence of false positives in certain regions. This non-uniform specificity may imply non-uniform sensitivity, since they are theoretically linked and likely share additional factors (e.g., nonuniform smoothness). Therefore, appropriately controlling the FWER may also result in relatively low sensitivity to true effects in many regions. In order to understand the extent of this issue, the proposed dissertation research is needed to quantify the spatial distribution of sensitivity and its relationship with specificity and smoothness.
The third aim describes a postdoctoral research direction to improve validity of fMRI methods. fMRI datasets are typically nested in several ways: e.g., regions within networks, runs within sessions. Multilevel models are important for such hierarchical data because they are better powered to reveal shared and unique features across the hierarchy. However, most fMRI analyses are conducted without models that explicitly account for the hierarchical nature of the data. Therefore, there is a need to identify appropriate multilevel measures, illustrate their utility, and implement them using popular softwares for the average user. In summary, this proposal will lead to the development of needed multilevel methods for fMRI, while enabling me to develop the training to become a leader in neuroimaging statistical methodology development.
Clinical neuroscience research using functional magnetic resonance imaging (fMRI) to inform individualized diagnosis and therapy is poised to flourish under major initiatives, but more accurate fMRI statistical methodology is needed. Recent work has raised questions about the reliability and validity major fMRI statistical methods that are both habitually underpowered and nonspecific. This proposal outlines a plan for ongoing dissertation research addressing these unresolved questions and for future postdoctoral research developing novel and better-powered multilevel fMRI methods that are more appropriate for the hierarchical nature of fMRI data.
