Gene-neuroimaging studies involve high-dimensional data that have a complex statistical structure and that are likely to be contaminated with outliers. Robust, outlier-resistant methods are an alternative to prior outliers removal, which is a difficult task under high-dimensional unsupervised settings. In this work, we consider robust regression and its application to neuroimaging through an example gene-neuroimaging study on a large cohort of 300 subjects. We use randomized brain parcellation to sample a set of adapted low-dimensional spatial models to analyse the data. Combining this approach with robust regression in an analysis method that we show is outperforming state-of-the-art neuroimaging analysis methods.
展开▼
