|
Abstract:
|
A common approach in associating brain images with phenotypes is to perform massively univariate analyses of individual voxels or vertices with a given phenotype. As with heritability in genetic analyses, a quantitative assessment of total variance explained by the entire image is of substantial interest in many applications. This is difficult to assess without assumptions, since the number of vertices or voxels can be much larger than the number of subjects, and brain structural measures are correlated spatially. We propose an estimator called BRAINWASH that consistently estimates the variance explained by the entire image under realistic conditions. This estimator is simple and easy to understand, and provides a basis for extensions such as “enrichment analyses” that can assess the relative importance of various regions or subnetworks of the brain. We demonstrate this method using a large multimodal brain imaging dataset from the ABCD study of almost 12,000 children.
|
